A mammogram can show dense tissue, suspicious spots, and, apparently, a tiny risk-prediction startup hiding inside the pixels.
That is the pitch behind a new study in Science Translational Medicine: instead of using mammograms only to ask, "Is there cancer right now?", researchers asked whether the image could also whisper, "Who is more likely to develop breast cancer over the next decade?" The answer, cautiously but loudly enough to hear over the espresso machine, was yes.
The paper, led by Mikael Eriksson and colleagues, tested a long-term, image-derived artificial intelligence model for breast cancer risk prediction in high-risk individuals. In plain English: the AI looked at mammograms and tried to estimate a person’s 10-year breast cancer risk. Not next Tuesday risk. Not "come back in six months and let’s panic politely" risk. Ten-year risk, which is the kind of timeline where prevention actually has room to stretch its legs. Eriksson et al., 2026
Your Mammogram Has Been Sitting on Data
Traditional breast cancer risk calculators use things like age, family history, prior biopsies, reproductive history, and breast density. Useful stuff. But also a bit like judging a whole movie from the trailer, the poster, and someone’s Letterboxd review.
A mammogram contains more. Texture. Asymmetry. Density patterns. Tiny structural weirdnesses too subtle for human eyes to turn into a neat checklist. Radiologists are excellent at finding current trouble, but they are not expected to mentally run a decade-long predictive model while also surviving clinic schedules designed by a printer with a grudge.
That is where AI can help. These systems can learn patterns across thousands or millions of images, then assign risk based on image features that may not map cleanly onto familiar categories. The machine is not "thinking" in the sci-fi sense. It is pattern matching at industrial scale, which is less glamorous but much more useful.
The New Product Launch: Prevention Mode
The study included individuals aged 31 to 94 from population-based cohorts in Olmsted County, Minnesota, and Sweden’s KARMA cohort, plus additional validation in the EMBED study in Atlanta. The median follow-up in the main cohorts was about 10 years.
The model performed well across settings. For invasive breast cancer, the 10-year time-dependent AUC was about 0.72 in both Olmsted and KARMA. AUC is a model-performance score where 0.5 is coin-flip territory and 1.0 is basically clairvoyance with a lab coat. So 0.72 is not magic, but it is meaningful. Think "promising beta release," not "ship it and fire all nuance."
The calibration also looked good: expected-to-observed ratios were close to 1.0, meaning the model’s predicted number of cancers roughly matched what actually happened. That matters because a risk model that ranks people well but wildly misestimates absolute risk is like a weather app that correctly says Tuesday is wetter than Wednesday but predicts 900 inches of rain. Technically directional. Practically unhinged.
Why the Top 10% Matters
One of the most useful findings was what happened when researchers looked at the highest-risk group. In KARMA, the AI tool identified about 33% of breast cancers within the top 10% of predicted risk. Traditional models did less well: Tyrer-Cuzick v8, BCSC v3, and Mirai captured about 23%, 20%, and 24%, respectively.
That is the heart of the story. If a health system can better identify a smaller group with a larger share of future cancers, it can offer more tailored prevention: supplemental screening, MRI for some people, risk-reducing medications for others, and closer conversations that do not sound like they were generated by a waiting-room pamphlet from 1998.
This fits a broader movement toward risk-based screening. The WISDOM randomized trial recently tested risk-based versus annual breast cancer screening and found risk-based screening was noninferior for stage at diagnosis, suggesting that personalization is not just a TED Talk with billing codes. Esserman et al., 2025
The Catch, Because Biology Always Has Terms of Service
AI risk prediction in mammography is not brand new. A 2024 systematic review found that image-only AI models often matched or beat traditional clinical risk tools, with median AUC around 0.72 compared with about 0.61 for density or clinical-factor models. Schopf et al., 2024
Other studies have shown similar momentum. Lehman and colleagues reported that deep learning outperformed traditional models for identifying women more likely to develop breast cancer. Lehman et al., 2022 External validation work has also tested mammography-derived AI models in diverse U.S. screening cohorts, including White and Black women. Gastounioti et al., 2022
But this field still has technical debt. Models need prospective testing, diverse populations, clear clinical workflows, and careful monitoring for bias. A tool that works beautifully in Sweden and Minnesota still needs to prove it can behave in the glorious chaos of real-world medicine, where patients miss appointments, scanners vary, insurance paperwork breeds in the walls, and biology refuses to read the product roadmap.
The Big Idea
This study nudges mammography from snapshot toward forecast. Not just "do we see cancer?" but "does this breast tissue pattern suggest elevated future risk?"
If reproduced and implemented carefully, that could shift prevention from broad averages to sharper targeting. Fewer people might get unnecessary extra testing. Higher-risk people might get earlier, more personalized options. Clinicians might finally have a tool that turns routine images into a prevention dashboard.
Cancer cells love ambiguity. This kind of AI tries to make the ambiguity smaller. That is not a cure. It is not a crystal ball. But it may be a smarter dashboard light, and in prevention, seeing the warning sooner can change the whole drive.
References
-
Eriksson M, Czene K, Scott C, Stoddard S, Smith H, Hall P, Vachon C. A long-term image-derived AI-based risk model for primary prevention of breast cancer in individuals at high risk. Science Translational Medicine. 2026;18(850). DOI: 10.1126/scitranslmed.ady7414
-
Schopf CM, Ramwala OA, Lowry KP, et al. Artificial Intelligence-Driven Mammography-Based Future Breast Cancer Risk Prediction: A Systematic Review. Journal of the American College of Radiology. 2024;21(2):319-328. DOI: 10.1016/j.jacr.2023.10.018. PMCID: PMC10926179
-
Lehman CD, Mercaldo S, Lamb LR, et al. Deep Learning vs Traditional Breast Cancer Risk Models to Support Risk-Based Mammography Screening. Journal of the National Cancer Institute. 2022;114(10):1355-1363. DOI: 10.1093/jnci/djac142
-
Gastounioti A, Eriksson M, Cohen EA, et al. External Validation of a Mammography-Derived AI-Based Risk Model in a U.S. Breast Cancer Screening Cohort of White and Black Women. Cancers. 2022;14(19):4803. DOI: 10.3390/cancers14194803. PMCID: PMC9564051
-
Esserman LJ, Fiscalini AS, Naeim A, et al. Risk-Based vs Annual Breast Cancer Screening: The WISDOM Randomized Clinical Trial. JAMA. 2025. DOI: 10.1001/jama.2025.24784
Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.